DeepSeek-V3.2 VRAM Requirements

Choosing a platform? Compare the 8×96GB, H200, B200/B300 and AMD Instinct hardware paths using the exact 689.483GB checkpoint boundary.

DeepSeek-V3.2 and DeepSeek-V3.2-Speciale are approximately 685B-parameter mixture-of-experts models, but parameter count alone is not the safest way to plan memory. Each official repository contains 163 safetensors files totaling 689,483,049,129 bytes, or 689.483GB in decimal units. The configs declare FP8 quantization with smaller BF16 and F32 tensors.

Short answer

640GB aggregate (8×80GB): the official files are 49.483GB larger than the nominal total. They do not fit wholly in accelerator memory before runtime overhead.
768GB aggregate (8×96GB): 78.517GB file-only margin. This is a lower-bound capacity check, not proof that an eight-device runtime has viable per-device shards or context headroom.
864GB aggregate: 174.517GB file-only margin.
1,024GB aggregate: 334.517GB file-only margin.

CPU or storage offload can make smaller accelerator configurations load, but placement and performance change. Report offload rather than calling it a resident fit.

The multi-GPU guide explains why aggregate memory is not one automatic pool. For smaller models, use the LLM VRAM calculator and hardware recommender.

Official repository facts

Official repository	Repository parameter metadata	Maximum positions	Safetensors files	Total bytes	Decimal GB
deepseek-ai/DeepSeek-V3.2	~685.355B tensors across FP8, BF16 and F32 metadata	163,840	163	689,483,049,129	689.483GB
deepseek-ai/DeepSeek-V3.2-Speciale	~685.397B tensors across FP8, BF16 and F32 metadata	163,840	163	689,483,049,129	689.483GB

Blob totals and tensor metadata come from the official Hugging Face repository APIs on October 25, 2026. Decimal GB matches repository file-size reporting. It is not GiB and excludes runtime allocations.

Both configs identify deepseek_v32, 61 layers, 256 routed experts, eight experts per token and max_position_embeddings: 163840. The repositories are not pure one-byte-per-parameter files: metadata reports mostly FP8 tensors plus smaller BF16 and F32 groups. That is why this guide sums the artifacts rather than multiplying a rounded parameter label.

File-only multi-accelerator boundaries

Nominal aggregate accelerator memory	Checkpoint difference	File-only result
512GB	177.483GB short	Does not fit by files
640GB (8×80GB)	49.483GB short	Does not fit by files
720GB	30.517GB margin	Files fit in aggregate; very constrained before runtime memory
768GB (8×96GB)	78.517GB margin	Files fit in aggregate; placement and headroom still unproven
864GB	174.517GB margin	Files fit in aggregate
960GB	270.517GB margin	Files fit in aggregate
1,024GB	334.517GB margin	Files fit in aggregate

Simple decimal subtraction from 689.483GB. Each device must still receive a viable shard plus runtime headroom; framework and interconnect support can rule out an aggregate total that looks sufficient.

Why eight 80GB GPUs are not enough by files

Eight nominal 80GB accelerators total 640GB, 49.483GB less than the official checkpoint files. No amount of tensor-parallel configuration turns 640GB into 689.483GB. A deployment can still load with host-memory or storage offload, but then it is not a wholly accelerator-resident eight-by-80GB run. The source should record how much resides on each device, host RAM, offloaded tensors, communication topology and observed transfer behavior.

Eight 96GB devices total 768GB and clear the file boundary by 78.517GB. That does not guarantee success: some allocations may be duplicated, shards may not balance evenly, expert placement can be constrained, and the framework needs room for cache, activations and communication buffers on every device.

Long context is a separate memory problem

The 163,840-position config value is an architecture ceiling, not a promise that the maximum sequence fits on a minimum file-capacity setup. KV cache, attention state, batch and concurrency consume additional memory. DeepSeek Sparse Attention is designed to reduce long-context computation, but “more efficient attention” does not eliminate checkpoint storage or prove a specific peak-memory figure.

For every result, state actual input length, generated tokens, batch, concurrency, cache precision and whether prefix caching or other runtime features were enabled.

V3.2 versus V3.2-Speciale

The official model card describes V3.2-Speciale as a high-compute reasoning variant. The two repositories currently have the same summed safetensors byte total, but they are different model artifacts and should not be treated as interchangeable benchmark workloads. Record the exact repository and revision. Output quality, reasoning behavior and speed are separate from file size.

Reproducible deployment checklist

Name the exact variant: DeepSeek-V3.2 or DeepSeek-V3.2-Speciale, plus revision or commit.
Record all devices: exact accelerator, form factor, exposed memory and count.
Describe topology: PCIe, NVLink/NVSwitch, Infinity Fabric or other fabric and host layout.
Record placement: tensor, pipeline, expert or data parallelism and per-device allocation.
Disclose offload: host RAM, CPU placement, storage-backed tensors and transfer behavior.
Pin software: runtime, version/commit, driver, kernels, operating system and launch flags.
Define the workload: input/output tokens, context setting, batch, concurrency and cache precision.
Measure separately: per-device peak memory, system memory, prompt processing, decode speed and time to first token.

Choose the next step

Plan multi-GPU placement · Compare NVIDIA data-center tiers · Compare AMD Instinct tiers · Check source-specific measurements · Build a reproducible record · Save private hardware criteria

Primary sources

Reviewed October 25, 2026. File-size arithmetic is not an independent benchmark. Model capability statements remain vendor claims unless separately reproduced.